H3 Progressive · Load Exact Frozen LOW Boundary (T8 EXP)

There's a point in a progressive workflow where you stop experimenting and start tuning: composition is locked, the small-canvas pass did its job, and everything left to try lives at full resolution. From that moment, the LOW half of your graph is dead weight. It eats VRAM, it eats queue time, and every time you accidentally hit Queue with the whole thing selected, it eats it again.

This node replaces that half with two strings.

What it does

In: artifact_path, artifact_sha256. Out: low_boundary, plan, report_json.

No MODEL loading. No LOW sampling. No assertion that today's controls match the frozen artifact - which sounds like a downside until you read it as honesty: the node knows what it can and can't prove, and it doesn't pretend. What it does prove is that the bytes you asked for are the bytes it loaded, and it hands you back the plan that the boundary was built with so the rest of the graph can trust its geometry.

It also validates. A corrupted, missing, or busy artifact fails, and it never silently regenerates. That last part is the one that matters: a node that says "hm, no boundary here, I'll just run the first pass for you" would be a nightmare in a graph where you deliberately removed the LOW model.

One implementation nicety you'll feel rather than see: the node's change fingerprint is computed from the artifact file itself, so if the file on disk changes (or you point at a different one), ComfyUI knows the node's output is stale and re-executes it properly. You don't have to nuke caches by hand.

Wiring the recovery graph

Load Exact Frozen LOW Boundary ─→ Boundary to External Lift Input → [learned 3D latent upscaler]
        (path + SHA)                 │                                        │
                                     └→ plan ────────────────→ HIGH Handoff Only → HIGH Sampler Only

Note what's missing: any LOW sampler, any LOW model loader, any LOW conditioning. That's the correct shape. If a stale LOW output node is still sitting in the graph and connected to a save or preview, remove it - the pack calls this out explicitly, and it's the difference between a cold graph and a warm graph wearing a cold graph's clothes.

Install

Manager, search MiniMax H3 Audio T8. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Restart ComfyUI completely, then refresh the page. The pack installs no Python packages by design - requirements.txt exists to keep installation from disturbing your Torch/CUDA - so nothing here can conflict with your environment. You still need current ComfyUI with native H3 support and the H3 model set for the HIGH side.

Where people get burned

Placeholder values. The Cold_HIGH example graphs in examples/workflows/44-progressive-split/ (and the resume_effects RF graphs, and the PDD cold graphs - this pattern is everywhere in the pack) ship with blank paths and hashes, because the author can't know yours. Queuing them unchanged gets you an error, and that error is the docs working as intended.

Typing the hash wrong by one character. It fails loudly, which is the correct behaviour and still annoying. Paste from the report.

And the tempting mistake: loading an old boundary to save time after changing your LOW settings. The load node won't stop you - it's doing exactly what you asked - but the result won't be the thing your controls describe, and the discrepancy is invisible in the output. The honest workflow is: frozen boundary means frozen inputs. Change something on the LOW side and you owe the graph a fresh LOW run and a fresh save.

One nice property of this design: the boundary artifact carries the evolving audio, not just video, so the HIGH pass keeps its joint AV clock even in a graph where LOW never ran. That's why a cold graph can produce a finished clip indistinguishable in structure from a warm one - the state it needs was all in the file.